global memory access disabled in a loop

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

i have a problem with the following kernel code :

for(int x = kernelradius; x <= nbNiveaux - kernelradius - 1; x++)

  {

	for(int y = kernelradius; y <= nbNiveaux - kernelradius - 1; y++)

	{

	  sommediv = 0;

				  unsigned int index_cmat = __umul24(i, nbNiveaux) + __umul24(__umul24(j, d), nbNiveaux) + x + __umul24(y , d); // (x,y)

	 float cmatxy = Cmats[index_cmat];

	 for(int kx = -kernelradius; kx <= kernelradius; kx++)

	 {

		for(int ky = -kernelradius; ky <= kernelradius; ky++)

		{

			dx = x + kx;

			dy = y + ky;

			unsigned int index_cmatf = __umul24(i, nbNiveaux) + __umul24(__umul24(j, d), nbNiveaux) + dx + __umul24(dy , d); // (dx,dy)

			CmatsF[index_cmatf] = h[kernelradius + kx + (kernelradius + ky) * largeurFiltre[0]];

		}

	 }	

	}

  }

CmatsF and h are global memory arrays. When I want to read elements of h outside the loops “for(int kx …, ky …” it works perfectly, but inside the loops, it seems to crash the kernel (or CUDA) and global memory is cleared. :unsure:

Can someone explain to me why I can’t access h in these loops ?

thanks

2

Broadly speaking, when trying to read global memory inside the loop, CUDA crashes. :wacko:

3

h[kernelradius + kx + (kernelradius + ky) * largeurFiltre[0]];

looks like is getting out of range, check the values on device emulation mode with a printf.